實驗以南海海域之 AIS 資料與其模擬之 RF 資料進行驗證。結果顯示，於 AIS 斷訊情境下，單一 AIS 必然全數遺失（0%），本系統將其還原為可確認之暗船以補足監控盲區，關聯覆蓋率達 100%，驗證融合管線之正確性；在多維規則消融實驗中，相較於單一條件判定，警報收斂率達 98% 以上；質化分析亦證實系統所篩選之高風險目標具備顯著異常之軌跡特徵與後續調查價值。本研究驗證了異質資料融合與多維物理規則整合，能在維持高可解釋性的前提下，同時改善暗船監控盲區與警報疲勞問題。;Maritime security is vital to global trade and national defense. However， current maritime surveillance relies heavily on the cooperative Automatic Identification System (AIS) and faces two practical bottlenecks. First， vessels engaged in smuggling or illicit operations often deliberately switch off their AIS (becoming so-called ``Dark Ships′′)， causing traditional single-source systems to lose track of the target and creating surveillance blind spots. Second， existing systems typically raise alerts using a single absolute physical threshold， which tends to flag legitimate vessels that slow down due to sea conditions， resulting in a large volume of alerts and alert fatigue.
To address these problems， this study proposes an automated maritime anomaly detection system that integrates heterogeneous data with multi-dimensional physical rules. The system is built upon a five-layer microservice architecture: when a target switches off its AIS， non-cooperative low-earth-orbit Radio Frequency (RF) signals are introduced， and a dynamic time-window interpolation is used to perform spatiotemporal correlation and trajectory-gap repair. Four physically interpretable behavioral features---speed drop， meandering， loitering， and dark-ship behavior in smuggling hotspots---are quantitatively modeled， normalized into standardized threat sub-scores via a Clamp function， and converged into a comprehensive threat score and risk level through a weighted-intersection model.
The system is evaluated using AIS data from the South China Sea together with RF data simulated from it. The results show that， under AIS-off conditions， the system improves the dark-ship identification coverage from 0% to 100%; in the multi-rule ablation experiment， the system achieves an alert convergence rate of over 98% compared with single-condition judgment; and the qualitative analysis confirms that the high-risk targets selected by the system exhibit significantly anomalous trajectory characteristics and investigative value. This study demonstrates that fusing heterogeneous data with integrated multi-dimensional physical rules can simultaneously mitigate dark-ship blind spots and alert fatigue while maintaining high interpretability.
